Accurate machine understanding of videos is important, e.g., to maintain the integrity and policy compliance of videos at a video hosting website. For example, a video-hosting website with a policy of hosting only family-friendly videos deploys video understanding systems to automatically exclude non-compliant videos. Techniques of image analysis are applied to individual video frames to understand the video. Due to the computational cost of processing individual frames, videos are sampled, e.g., at a rate of one frame per second, and only the sampled frames are subjected to image analysis.
Sampling-based video understanding is susceptible to attack. For example, such systems fail to detect a non-compliant video, if the video includes policy-compliant frames inserted to match sampling instants. This disclosure utilizes randomization of sampling instants to thwart frame insertion attacks that attempt to mask actual video content. Randomized sampling for video understanding also assures reproducibility such that the understanding of the video is independent of the sampling instants.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Lee, Juhyun, "Non-deterministic video frame sampling to thwart frame insertion attacks", Technical Disclosure Commons, (September 20, 2017)